Nothing in the review loop names a provider. It asks for a purpose; the purpose resolves to a logical model; the logical model is the only thing that knows which connection and which model id get called.
Two objects, and the difference matters:
- An AI purpose is a fixed slot the review loop asks for - review generation, triage, verification, embeddings. The set is fixed; you cannot add one.
- A logical model is a name you choose (
fast,deep,embed- anything) that points at one model on one connection and carries that model’s reasoning effort and protocol mode.
That indirection is what you get for the extra object: because no client’s review configuration names a provider, you can move a tenant from OpenAI to Anthropic by repointing the names, and no purpose mapping, review pass or prompt override has to change.
Everything operational lives elsewhere. Mapping purposes to names, what happens when one resolves to nothing, and where reasoning effort and protocol mode are set are all in purposes, effort and protocol. Which providers a name can point at is in AI providers, and getting models attached to a connection in the first place is in models, the catalog and prices.